Road Crack Detection Of 3D Ground Penetrating Radar Image Based On Deep Learning

With the vigorous development of infrastructure projects,the number of roads in China is increasing rapidly year by year.However,with the increase of driving pressure and the damage caused by natural factors,roads will inevitably produce various diseases,which are easy to cause traffic accidents and threaten driving safety.The blockade and maintenance of roads have also hindered economic development.As one of the most common road distress,cracks are usually the root cause of other serious distresses.If they are not detected and repaired in time,the cracks will damage the subgrade,and then evolve into loose,void and other diseases,and even cause road collapse.Therefore,road cracks detection is particularly important,but the traditional manual detection methods are usually inefficient,low accuracy,time-consuming and labor-consuming.What’s worse,some of the cracks are hidden so that the conditions of cracks in the road can’t be observed by eyes.Therefore,in order to detect and repair the road cracks in time and avoid more serious road disasters,based on the existing research results,this paper realizes the automatic detection of road cracks by using deep learning algorithm and 3D ground penetrating radar,which has certain engineering significance.The main work of this paper is as follows:In this paper,the deep learning algorithm is used to detect the cracks in 3D ground-penetrating radar images.The detection process includes two steps: using classification network to filter the slab images and using semantic segmentation network to segment the cracks in radar images.In the process of road crack detection,aiming at the problem that the semantic segmentation network is easy to identify the edge of the slab as cracks,this paper uses the improved network based on Efficient Net to train a slab classifier to eliminate these images.At the same time,the characteristics of Efficient Net alleviate the problem of low efficiency caused by the serial use of the two network.The test results in the test set show that the accuracy of the classifier reaches 97%.In the process of image feature extraction by semantic segmentation network and in view of the difficulty of crack feature extraction caused by slender and irregular crack shape and small proportion in the image,a Parallel Multi-Directional Attention mechanism(PMDA)block is proposed to establish an association relationship between a single pixel of the feature and its adjacent pixels,so as to give higher weight to the pixels of the crack feature,which can enhance the ability of crack feature extraction of the network.Aiming at the problem that the various working conditions and the interference of radar electromagnetic wave lead to various noises in the radar image,which reduces the effectiveness of the features extracted by the network,this paper proposes a Multi Convolution-Feature Selection(MC-FS)block to extract features in parallel in order to increase the diversity of features.Finally,the channel attention mechanism is used to weight the channel,so as to realize the transformation from diversity to effectiveness.At the same time,the Residual Squeeze(RS)block is designed to compress the number of the features’ channels,reduce network parameters and improve network computing efficiency.Finally,this paper uses the above three modules to improve UNet and forms a semantic segmentation network named Crack UNet,and uses the network to realize the crack segmentation of radar images.The test results in the test set show that the MPA and MIo U of Crack UNet reach 85.69% and 75.02% respectively.The detection accuracy at the image level has also reached 93%,which has achieved better performance compared with the other semantic segmentation network.Ablation Experiment and attention block contrast experiment also proved the effectiveness of the above three modules.Finally,this paper encapsulates slab classifier and semantic segmentation network into crack detection module,and combines data generation module and joint judgment module to form an automatic road disease detection system.